Role-Based Decision Authority Taxonomy

• The topic defines decision authority as the formal right to make final decisions, differentiating roles such as Decision Support, Contributor Intervention, and Full Autonomy.
• It employs a 20-level taxonomy within the human-AI task tensor, partitioning authority into human-centric and AI-centric configurations with fixed and emergent paradigms.
• Industrial examples like smart grids and autonomous drilling demonstrate its practical application in optimizing control and enhancing system efficiency.

A role-based taxonomy of decision authority provides a structured framework to specify, compare, and analyze the assignment and exercise of decision rights within human-AI collaborations and multi-agent systems. This taxonomy serves to formalize not only who holds the decisive power in a given sociotechnical system but also how the interplay of oversight, delegation, and autonomy follows from specific organizational designs in both generative AI work and hierarchical multi-agent architectures (Doshi et al., 6 Jan 2025, Moore, 18 Aug 2025).

1. Formal Definitions and Theoretical Foundations

Decision authority is formally characterized as the allocation of the formal right to make a final decision (the Decision-Maker), as well as delineation of the role for any secondary agent (the Contributor) regarding support, oversight, or notification in the process (Doshi et al., 6 Jan 2025). In hierarchical multi-agent systems, authority is primarily governed by the "Role and Task Delegation" axis, which spans a spectrum from fixed, pre-defined roles to emergent, dynamically allocated roles (Moore, 18 Aug 2025). In fixed-role paradigms, an agent’s authority is static and codified a priori, whereas in emergent-role systems, authority may transition among agents via negotiation, learning, or dynamic election.

In the context of the human-AI task tensor, decision authority ($D$) is a discrete dimension, with 20 mirrored levels identifying both human-centric and AI-centric variants of decision rights. Formally, a task can be denoted as:

$$T = (\mathrm{TaskDef}, \mathrm{AIInt}, \mathrm{Modality}, \mathrm{AuditReq}, \mathrm{OutputDef}, D, \mathrm{AIType}, \mathrm{UserType})$$

with $$D \in \{\mathrm{H}1,\ldots,\mathrm{H}10, \mathrm{A}1,\ldots,\mathrm{A}10\}$$ specifying the decision authority configuration (Doshi et al., 6 Jan 2025).

2. Taxonomic Structure: Levels and Categories of Authority

The 20-level decision authority spectrum in the human-AI task tensor is partitioned into two parallel sets: ten cases with the human as Decision-Maker (H1–H10), and ten with AI as Decision-Maker (A1–A10). Each set is structured into three subcategories: Decision Support, Contributor Interventions, and Full Autonomy.

Subcategory	Level	Human DM (Contributor=AI)	AI DM (Contributor=Human)
Decision Support	1	H1: AI provides many suggestions	A1: Human provides many examples
	2	H2: AI provides a few suggestions	A2: Human provides a few examples
	3	H3: AI provides a single suggestion	A3: Human provides a single example
Contributor Interventions	4	H4: Human must approve AI proposal	A4: AI must approve human proposal
	5	H5: Human can veto AI decision	A5: AI can veto human decision
	6	H6: Human observes AI decisions	A6: Human observes AI decisions
	7	H7: AI requests human input if needed	A7: Human requests AI input if needed
	8	H8: AI informs human at its discretion	A8: Human informs AI at discretion
Full Autonomy	9	H9: Human decides without informing AI	A9: AI decides without informing human
	10	H10: Human only (no AI involvement)	A10: AI only (no human involvement)

Decision Support levels involve the Contributor serving as an advisor. Contributor Intervention stages mark a gradient from explicit approval and veto to observation and selective informing. Full Autonomy levels are characterized by the absence of the Contributor's role or awareness (Doshi et al., 6 Jan 2025).

3. Role Classes and Authority Scopes in Multi-Agent Systems

Within hierarchical multi-agent systems, decision authority is inherently role-based and tied to distinct agent classes, each with specific scopes and constraints of control (Moore, 18 Aug 2025):

• Strategic Manager / Coordinator: Sets global objectives, creates and delegates tasks, aggregates information, but does not execute primitive actions.
• Submanager / Aggregator: Operates at a tactical level, translating strategic goals into subgoals and aggregating operational feedback.
• Operational Executor / Worker: Executes assigned subtasks within defined bounds, potentially negotiating with peers if roles are emergent.
• Symmetric Peer / Consensus Participant: Participates in distributed decision-making processes, with authority emerging from local interactions rather than hierarchy.
• Auctioneer / Market Mediator: Allocates resources based on bids/asks, constrained by market protocols.
• Emergent Leader / Dynamic Role Holder: Temporarily holds elevated decision rights, arising from runtime election, bidding, or negotiation.

The assignment of these roles, their authority scopes, and the transitions between them are mediated by both system design and runtime protocols (e.g., contract-net, auctions, learning-based delegation).

4. Decision Authority: Interaction with Organizational and System Axes

Decision authority is intertwined with other dimensions of both generative AI task frameworks and multi-agent system taxonomies:

• Control Hierarchy: Centralized designs yield static, top-weighted authority (e.g., single manager), while decentralized or hybrid structures permit emergent roles and distributed authority.
• Information Flow: Top-down flow enforces managerial authority; bottom-up flow enables situational awareness for decision-makers; peer-to-peer exchanges underpin distributed and emergent modes.
• Temporal Layering: Authority is scoped by temporal horizons—strategic managers operate at long time constants; operational executors handle rapid decisions at short horizons.
• Communication Structure: Static (tree/star) topologies enforce fixed roles; dynamic networks enable role flexibility and emergent delegation protocols. Protocols such as auctioneering require specific connectivity (star around broker) for effective decision mediation (Moore, 18 Aug 2025).

In the AI task tensor, relationships are further detailed: as $D$ transitions from H1 to H10, AI integration typically increases from complementary to substitutive; high audit requirements cluster around human-centric authority, while well-defined output tasks allow for greater AI autonomy (Doshi et al., 6 Jan 2025).

5. Industrial and Applied Examples

Concrete applications illustrate the instantiation of role-based decision authority:

• Smart Grids: Device-level agents hold operational authority over local set-points; microgrid-level agents exercise tactical authority over intermediate signals; main grid agents maintain strategic authority for dispatch targets, configured as a fixed three-tier hierarchy (Moore, 18 Aug 2025).
• Oilfield Operations: Sensor agents autonomously trigger safety actions, intervention agents schedule maintenance, and field operations agents set aggregate production—all mirroring human organizational design but with agent roles (Moore, 18 Aug 2025).
• Autonomous Drilling with Contract Net: Manager agents delegate by broadcasting subgoals; workers bid for task execution, with successful bidders acquiring operational authority for the subtask; the authority structure is episodic and partially emergent (Moore, 18 Aug 2025).
• Human-AI Collaboration in Generative AI: Levels such as H2 (AI suggesting options for human selection) or A5 (AI executing unless vetoed by a human compliance officer) directly reflect the granularity of allocation and oversight possible with the 20-level taxonomy (Doshi et al., 6 Jan 2025).

6. Role-Based Taxonomy: Schematic Synthesis

A synthesized schema from the hierarchical MAS taxonomy aligns role, authority, information flow, communication topology, temporal scope, and fixity:

Role	Authority Scope	Temporal Layer	Fixed vs Emergent
Strategic Manager	Global objectives, task creation	Long-horizon	Fixed
Tactical Coordinator	Subgoal translation, scheduling	Mid-horizon	Usually Fixed
Operational Executor	Primitive action execution	Short-horizon	Fixed
Auctioneer / Broker	Resource allocation, pricing	Round-based	Fixed per auction
Symmetric Peer	Collaborative agreement	Single layer	Emergent
Emergent Leader	Temporary task leadership	Episodic	Emergent

This table encapsulates the systematic mapping of agent roles to their decision authority characteristics within hierarchical and distributed system designs (Moore, 18 Aug 2025).

7. Significance and Analytical Utility

The role-based taxonomy of decision authority enables precise specification, analysis, and engineering of mixed human-AI and multi-agent systems. By distinguishing between fine granularity of oversight, intervention, and autonomy, it informs both organizational design and technical system deployment. The human-AI task tensor's decision authority axis underpins analytical tractability for research on augmentation versus automation, aligns authority with audit and output structures, and serves as a reference for future systems combining LLMs, policy agents, and human experts (Doshi et al., 6 Jan 2025). In multi-agent contexts, explicit formalization of roles and their associated authority supports rigorous evaluation of system efficiency, robustness, and scalability in industrial, critical infrastructure, and novel AI-managed domains (Moore, 18 Aug 2025).